CAHIER SCIENTIFIQUE2021S-07 CS
TOUHAMI ABDELKHALEK
DOROTHÉE BOCCANFUSO
Impact of Tax Reforms in Applied Models: Which Functional Forms Should Be Chosen for the Demand System? Theory and Application for Morocco
The purpose of the Working Papers is to disseminate the results of research conducted by CIRANO research members in order to solicit exchanges and comments. These reports are written in the style of scientific publications. The ideas and opinions expressed in these documents are solely those of the authors.
Les cahiers de la série scientifique visent à rendre accessibles les résultats des recherches effectuées par des chercheurs membres du CIRANO afin de susciter échanges et commentaires. Ces cahiers sont rédigés dans le style des publications scientifiques et n’engagent que leurs auteurs.
CIRANO is a private non-profit organization incorporated under the Quebec Companies Act. Its infrastructure and research activities are funded through fees paid by member organizations, an infrastructure grant from the government of Quebec, and grants and research mandates obtained by its research teams.
Le CIRANO est un organisme sans but lucratif constitué en vertu de la Loi des compagnies du Québec. Le financement de son infrastructure et de ses activités de recherche provient des cotisations de ses organisations-membres, d’une subvention d’infrastructure du gouvernement du Québec, de même que des subventions et mandats obtenus par ses équipes de recherche.
CIRANO Partners – Les partenaires du CIRANO Corporate Partners – Partenaires corporatifs Autorité des marchés financiers Bank of Canada Bell Canada BMO Financial Group Business Development Bank of Canada Caisse de dépôt et placement du Québec Desjardins Group Énergir Hydro-Québec Innovation, Science and Economic Development Canada Intact Financial Corporation Manulife Canada Ministère de l'Économie, de la Science et de l'Innovation Ministère des finances du Québec National Bank of Canada Power Corporation of Canada PSP Investments Rio Tinto Ville de Montréal Academic Partners – Partenaires universitaires Concordia University École de technologie supérieure École nationale d’administration publique HEC Montréal McGill University National Institute for Scientific Research Polytechnique Montréal Université de Montréal Université de Sherbrooke Université du Québec Université du Québec à Montréal Université Laval CIRANO collaborates with many centers and university research chairs; list available on its website. Le CIRANO collabore avec de nombreux centres et chaires de recherche universitaires dont on peut consulter la liste sur son site web.
© February 2021. Touhami Abdelkhalek, Dorothée Boccanfuso. All rights reserved. Tous droits réservés. Short sections may be quoted without explicit permission, if full credit, including © notice, is given to the source. Reproduction partielle permise avec citation du document source, incluant la notice ©.
The observations and viewpoints expressed in this publication are the sole responsibility of the authors; they do not necessarily represent the positions of CIRANO or its partners. Les idées et les opinions émises dans cette publication sont sous l’unique responsabilité des auteurs et ne représentent pas nécessairement les positions du CIRANO ou de ses partenaires.
ISSN 2292-0838 (online version)
Impact of Tax Reforms in Applied Models: Which Functional Forms Should Be Chosen for the Demand
System? Theory and Application for Morocco *
Touhami Abdelkhalek †, Dorothée Boccanfuso ‡
Abstract/Résumé
When researchers and policymakers conduct impact analyses of economic reforms, especially fiscal reforms, the specification of the household demand system becomes crucial. There is a trade-off between using demand systems simple to manipulate but less realistic and other systems that are more realistic but often more complex and difficult to estimate or calibrate. In this paper, we compare the results from two different demand systems: a simple one, the Cobb-Douglas (CD), and a more complex one, the Constant Difference Elasticity (CDE). We develop an hybrid method of estimation - calibration based on the estimation of the parameters and elasticities of a QUAIDS system and on the calibration of those of the CDE system using a cross-entropy approach. The estimates obtained are introduced into a micro-simulated partial equilibrium model to approximate the impact of the VAT reform on poverty measures in Morocco. We show that when the simulated shocks are moderate, the gain of using a CDE system instead of a CD system is marginal but, when these shocks are stronger, the differences become significant and increase. Then the use of these models can lead to different results when evaluating public policies and their impacts on poverty measures.
Keywords/Mots-clés: Demand Systems, Estimation-Calibration, Tax Reform, Morocco
JEL Codes/Codes JEL: C51, D12, I32, H31
* Author for correspondence: [email protected]† Institut National de Statistique et d'Economie Appliquée - Morocco‡ Université Mohammed VI Polytechnique - FGSES - Morocco, CIRANO - Montréal
1 Introduction and objective
In a context of general economic reforms in several countries in the Middle East and North
Africa (MENA), including Morocco, tax reforms are often put on the agenda. During the
3emes National Assises on Taxation held in May 2019 in Skhirat, under the theme ”tax eq-
uity”, a working group was formed to explore the neutrality of value added tax (VAT), make
a diagnosis on the purchasing power of consumers and propose reform scenarios. Researchers,
government authorities and international organizations try to approach the impacts of such
reforms on the economy as a whole and on households in particular. Indeed, the latter are
directly impacted following the implementation or after the changes in the VAT rates that
directly or indirectly, modify the price levels of various goods and services. These reforms
naturally change the levels of monetary poverty measures and household welfare measures
(see for example, De Quatrebarbes et al. (2016), Ndemezo and Baye (2016) or Boeters et al.
(2010)).
In terms of the economic analysis tools used to anticipate and assess these various impacts,
several models are usually used. These are computable general equilibrium (CGE) models
or micro-simulated calculable partial equilibrium models such as the one we propose in this
article. These models use household survey data. At the level of household utility and/or de-
mand functions used in these models, several functional forms are proposed in the literature.
Each has its pros and cons. Some are rather simple to use with parameters that are easy
to estimate or calibrate but often have a priori unrealistic properties. Others are, on the
other hand, much more flexible and presumably, better approximate household behaviour
but reveal parameters that are difficult to estimate and even to calibrate. The choice be-
tween these two types of specifications often raises problems for researchers, who are faced
with a dilemma between facility, complexity, rigour, reliability and credibility of the results.
In this paper, we propose to approach and empirically examine the gap between two demand
systems. The first is the simplest and derives from the well known Cobb-Douglas direct utility
function (CD). The second is more flexible but more complex; it is the one resulting from
1
the Constant Difference Elasticity (CDE). This demand system was introduced by Hanoch
(1975).
At the methodological level, we develop an hybrid estimation-calibration approach that
involves several steps to approximate the parameters of the CDE demand system based on
Yu et al. (2004), Chen (2017) and van der Mensbrugghe (2020). This method consists of
calibrating the elasticities and parameters of the CDE system using estimates obtained from
another demand system (AIDS1 or QUAIDS2, for example) and minimizing an appropriate
entropy family criterion.
This article contains three contributions. The first one is theoretical. It aims to improve the
estimation-calibration of the parameters of the CDE demand systems by clearly proposing
for the first time to our knowledge, the necessary econometric and numerical steps to link the
elasticities of AIDS or QUAIDS with those of the CDE. The second consists in estimating-
calibrating the parameters and elasticities of a CDE system on Moroccan data under an
appropriate aggregation using the proposed approach, which is a significant added value in
itself. The third contribution is clearly in terms of economic policy approach. Indeed, our
contribution allows to compare the results in terms of poverty of a tax reform (VAT reform in
Morocco) under two alternative expenditure systems (CD and CDE). This third contribution
is of major interest to compare and choose the specification to be retained in different
economic policy models at the level of the household demand module, in microsimulated
models in particular.
This article is articulated as follows. In the section 2, we begin with a brief literature
review of the theoretical framework for demand systems in applied models derived. In the
subsection 2.1, we briefly present the classic demand system generated from a Cobb-Douglas
(CD) utility function. We then present in more detail the Constant Difference Elasticity
(CDE) demand system and its properties (Subsection 2.2). As our approach is based on
the estimation of the Quadratic Almost Ideal Demand System (QUAIDS) to calibrate the
1Almost Ideal Demand System (AIDS)2Quadratic Almost Ideal Demand System (QUAIDS)
2
elasticities and parameters of the CDE system, we first present this demand system in the
subsection 2.2.1. We formally deduce the demand equations and then the elasticities of
the latter. We also recall the major microeconomic conditions that these elasticities must
respect.
In section 3, our hybrid estimation - calibration method is developed and we present the
data used. We detail the methodology to proceed from the estimation of a QUAIDS de-
mand system to the calibration of the parameters and elasticities of a CDE demand system.
More specifically, we present the canonical form of the QUAIDS model and its extensions,
in particular the one that integrates demographic variables. We then develop the calibration
approach based on the minimization of a cross-entropy criterion. Finally, we present the
algorithm we developed to calibrate the elasticities and the parameters of the CDE demand
system using GAMS software. The subsection 3.2 presents the 2019 wave of the Enquete
Panel des Menages (EPM) of the Observatoire national du developpement humain (ONDH)
of Morocco which is the main data used for the estimations. We then dedicate the subsection
3.3 to the implementation of the proposed method by breaking it down into six steps that
allow to move from theory to practice.
In the next section (Section 4), we present the twelve (12) scenarios considered for simu-
lations according to the variation of household expenditures and taxation systems (VAT).
Comparative point estimation and stochastic-dominant analyses are conducted to highlight
the impact of the reforms depending on the demand systems chosen. In the last section
(Section 5), we conclude and make recommendations to the researchers who use this type of
model and to the policy makers who use the results produced by these researchers.
3
2 Theoretical framework : from the microeconomics
to the applied models
In applied models, it is necessary to assign a value for the parameters of some behavioural
functions. Some parameters are sometimes estimated on an ad hoc basis or can be taken
from applications in other countries requiring situations to be comparable (Abdelkhalek and
Dufour, 1998).
In general, parameter calibration and estimation are crutial in applied models as they drive
the results. On the consumption side, the household expenditure module is critical. Its
calibration is therefore fundamental for simulating the impact of fiscal and public policies.
Since one want to capture the transmission mechanisms of the effects of these policies on
households, it is interesting and even crucial to introduce heterogeneity between them. This
heterogeneity can be introduced based on household budget shares and assuming a rich and
flexible demand system that will allow for enhanced microsimulation analysis (Savard, 2004).
In practice, some of these constraints are not respected in certain demand systems. It is
precisely at this level that the trade-off is made in favour of simple but probably unrealistic
models. In the next section, we present the two demand systems on which we base our
comparison, derived from a CD utility function and the CDE demand system.
2.1 Demand system based on the Cobb-Douglas utility function
The Cobb Douglas (CD) utility function is a special case of the Constant Elasticity of
Substitution (CES) utility function (when elasticity of substitution equal to one) and of the
Linear Expenditure System (LES) utility function (when the minima are zero) (Shoven and
Whalley, 1992).
Formally, a CD utility function is as follows 3:
3The properties are presented for example in Shoven and Whalley (1992)
4
u(q) =k∏i=1
qαii (1)
where qi is the consumed quantity of good i and αi > 0. Without loss of generality, we often
impose∑k
i=1 αi = 1.
The Marshallian demand functions are as follows4:
qi =αi∑kj=1 αj
.m
pi=αim
pi∀i = 1, . . . , k (2)
where m is the total household expenditure and pi, the price of good i. From this de-
mand system, several properties are derived. Those we are particularly interested in are the
following:
• αi = wi, the budget share of the good i in the total expenditure and is constant for all
m and pi ;
• The elasticity of total expenditure, noted ηi is unitary for any i product;
• The uncompensated own price elasticity, denoted σuii is equal to −1 for every product
i;
• The compensated own price elasticity, denoted σcii is equal to −(1 − αi) for every
product i;
• The demand for each good i depends only on the price of the good i and not on the
price of other goods. This implies that the cross-price elasticity, denoted σuij is null
∀i 6= j.
This demand system also respects all the basic microeconomic properties that we summarize
in subsection 2.2.2.
The problem with this relatively simple demand system is that it seems to be limited when
assessing the impact on the demand for goods and services as a result of changes in economic
4To simplify the notation, we have not indexed household variables and parameters with h.
5
policy (Yu et al., 2004). For example, with this demand system derived from CD utility
function, as mentioned previously, income and price elasticities are constant and then do not
depend on budget shares. Because of these characteristics, the demand system is considered
unrealistic and inflexible (see for example, Sadoulet and De Janvry (1995), Rimmer and
Powell (1996) and Corong et al. (2017)).
To overcome at least this limit, several other more flexible and apparently more realistic
demand systems have been introduced in this literature. These are the Almost Ideal Demand
System (AIDS) family proposed by Deaton and Muellbauer (1980a) and the CDE system
that we discuss in the next section.
Deaton and Muellbauer (1980a) developed the AIDS demand system, based on an approx-
imation of the first-order conditions of a demand system. Even this one has interesting
properties both from a theoretical point of view and in its estimation, it does not capture
the non-linearities of the Engel curves, even though they are empirically observed. By in-
troducing a quadratic term in the AIDS specification, Banks et al. (1997) propose a more
flexible demand system that respects the Engel condition and other constraints of the AIDS
system.
In this literature on demand systems, already in 1975, Hanoch introduced the functional
form called Constant Difference of Elasticity (CDE). It lies between the CES form and more
flexible specification. The CDE demand system allows richer representation of the effects of
income (Dimaranan et al., 2006). Because of its properties and the number of parameters
(only three for each good), the functional form of the CDE offers many potential applications
(See Surry (1993), Yu et al. (2004) and van der Mensbrugghe (2020)).
However, it seems that few researchers have directly estimated the CDE parameters because
of the complexity to do so. The few examples that exist in terms of estimation are exclu-
sively on the production side or based on cost functions (Surry, 1993). Yu et al. (2004),
Chen (2017) and van der Mensbrugghe (2020) propose a hybrid estimation-calibration ap-
proach that involves several steps. This method consists of approaching the elasticities and
6
the parameters of the CDE demand system using estimates from another demand system
(e.g. AIDS, AIDADS, QUAIDS, ˙...) and minimizing a criterion, respecting only some mi-
croeconomics conditions of a demand system. This is especially relevant since it has been
shown that the CDE system can violate some of Engel’s properties in its empirical version
(Yu et al., 2004). Considering this, we improve these approaches taking into consideration
a larger number of microeconomic constraints than what is done by Chen (2017), Yu et al.
(2004) and van der Mensbrugghe (2020) as we will see in the next section.
2.2 A Constant Difference Elasticity demand system
In this article, we compare the impact of several tax reforms on poverty measures using
two demand systems: the first one generated from a simple CD utility function and the
second one from the more complex, CDE demand system. The calibration of the elasticities
and parameters of this system necessarily requires the estimation of the parameters and
elasticities of the AIDS or QUAIDS demand system.5 Consequently, we describe this demand
system with some detail before presenting the CDE demand system.
2.2.1 The Quadratic Almost Ideal Demand System - QUAIDS
Using the notation of Banks et al. (1997), the functional form of the QUAIDS for a household
is as follows:
wi = αi +k∑j=1
γij ln pj + βi ln
{m
a(p)
}+
λib(p)
[ln
{m
a(p)
}]2, i = 1, . . . , k (3)
where wi is the budget share of the good i defined as :
wi =piqim
, (4)
5At this step, other flexible demand systems could be used for the estimation of the parameters andelasticities.
7
m the total household expenditure, pi the price of good i and qi, the Hicksian demand for
good i by the household, k is the number of considered goods.
By definition:
ln a(p) = α0 +k∑i=1
αi ln pi +1
2
k∑i=1
k∑j=1
γij ln pi ln pj (5)
is a translog price index and
b(p) =k∏i=1
pβii (6)
a price index. This system must respect a set of microeconomic conditions referring to
additivity, homogeneity and symmetry that are expressed in constraints on parameters for
each household (αi, βi, γij and λi):6
k∑i=1
αi = 1,k∑i=1
βi = 0,k∑j=1
γij = 0,k∑i=1
λi = 0, and γij = γji. (7)
Note λi, αi, βi and γij are the parameters of the model QUAIDS that will be estimated.7
In addition, control variables relating to household characteristics are introduced to the
above equations to avoid specification errors. In particular, these are demographic variables
that characterize either the household or specific members of the household, such as the head
of the household. The variables encountered in this context are: household size, household
size squared, number of children, number of adults, gender, age of the head of household,
socio-professional category, level of education. Furthermore, variables regarding the region
or place of living of the household can also be introduced as explanotary variables if they
were not used as stratification variables in the estimation itself.
Based on Ray (1983), Poi (2012) notes that modeling heterogeneity in this way does not mean
that the heterogeneity enters the model in a linear fashion. As we saw previously, hetero-
geneity appears not only linearly in the intercepts, but also non-linearly in all expenditures
6When λi = 0, ∀i = 1, . . . , k, in Equation 3 (this can be tested after parameters estimation) thequadratic term disappears and the QUAIDS becomes AIDS system, respecting the conditions of symmetry,separability and Cournot and Engel.
7See Diagram 1.
8
through the first price aggregator. Following Poi (2012), the expression (3) becomes:8
wi = αi +k∑j=1
γij ln pj + (βi + ν ′iz) ln
{m
m0(z)a(p)
}+
λib(p)c(p, z)
[ln
{m
m0(z)a(p)
}]2(8)
with z the vector of household socio-demographic characteristics, and
m0(z) = 1 + ρ′z (9)
c(p, z) =k∏j=1
pν′jz
j (10)
where νj represents the j th column of s× k parameter matrix ν with
k∑j=1
νrj = 0 for r = 1, . . . , s (11)
the additivity condition.
Another common problem in estimating demand systems is that of endogeneity in both
total expenditures and prices. As noted by Deaton and Muellbauer (1980a), Attfield (1985)
and many other authors after, the simultaneous nature of the demand system is one of the
reasons for this endogeneity. Indeed, since the total expenditure is defined as the sum of the
expenses for each product, and since these expenditures are assumed to be endogenous, it
can be assumed that the total expenditure is jointly endogenous and thus correlated with
the error term. Therefore, if expenditures or prices are correlated with errors, the resulting
estimators are biased and inconsistent.
For the price variables, when regionally and goods-specific prices do not exist, it is common in
empirical studies to determine prices by relating the value spent on a good to the consumed
quantity if available. However, differences in prices observed between households may reflect
8Here again, if we set H0 : λi = 0, ∀i = 1, . . . , k, we are left with the AIDS model with demographicsused by Ray (1983). This null hypothesis can be tested.
9
something other than the market price and may, for example, differ because of a difference
in quality (Deaton, 1988).
As usual, this problem of endogeneity can be addressed by using instrumental variables.
However, if the demand system is large with a large number of goods, we will need at least
as many instruments as endogenous variables. This is one of the limitations of this approach
in practice. This is, moreover, the main reason why many authors do not take into account
the problem of endogeneity, particularly concerning prices in the demand system estimation
context.
In this article, as explained in the data section 3.2, we sought to have prices differentiated by
region and location to attempt to avoid endogeneity by determining prices directly from the
information in the survey. Furthermore, we tested for the exogeneity of the total expenditure
but not for the exogeneity of the prices which were constructed as explained in section 3.2.
2.2.2 The Constant Difference Elasticity - CDE
In this literature on demand systems, Hanoch (1975) introduced the functional form called
Constant Difference Elasticity (CDE). Following her and Chen (2017), let M be an expendi-
ture function with a price vector p associated to a Hicksian demand system noted q. Formally,
m0 = M (p0, u) ≡ {min p0q0 : f (q0) ≥ u} where the 0 index refers to the base situation. By
normalization, we postulate M(
pm0, u)
=∑
i βiuei(1−αi)
(pim0
)1−αi
≡ 1.
Using Roy’s identity and the implicit function theorem, they deduce the associated Hicksian
demand functions for a CDE demand system:
qi =βiu
ei(1−αi) (1− αi)(pim0
)−αi
∑j βju
ej(1−αj) (1− αj)(pjm0
)1−αj(12)
with qi the Hicksian demand for good i by the household, pi the price, αi a substitution
parameter of the household expenditure function, βi a scale parameter and ei an expansion
10
parameter with βi > 0, ei > 0 and 0 < αi < 1.9
Let σij be Allen-Uzawa elasticity of substitution (AUES) obtained from the CDE demand
system10:
σij = αi + αj −∑i
wiαi −δijαiwi
(13)
with wi the share of household expenditure in good i, δi,j = 1 when i = j and 0 else. Thus,
the own-AUES for the good i is equal to:
σii = 2αi −∑i
wiαi −αiwi
when i = j (14)
and the cross-AUES:
σij = αi + αj −∑i
wiαi when i 6= j. (15)
Following Hanoch (1975), the uncompensated demand elasticities for the CDE demand sys-
tem are11:
σuij = wj
[αj −
ei(1− αi)∑iwiei
−∑
iwieiαi∑iwiei
]− δijαj (16)
where δi,j = 1 when i = j and 0 else such that:
σuii = wi
[αi −
ei(1− αi)∑iwiei
−∑
iwieiαi∑iwiei
]− αi when i = j (17)
and
σuij = wj
[αj −
ei(1− αi)∑iwiei
−∑
iwieiαi∑iwiei
]when i 6= j. (18)
Furthermore, following Hanoch (1975)12, the household expenditure elasticities are equal to:
ηi =
(∑i
wiei
)−1 [ei (1− αi) +
∑i
wieiαi
]+
(αi −
∑i
wiαi
). (19)
9For details and notations see Hanoch (1975) and Chen (2017).10See equation 3 page 409 in Hanoch (1975).11See page 414 in Hanoch (1975).12See equations 3.21 and 3.22 page 413 in Hanoch (1975).
11
Note that wi, ηi, σuij and σcij respectively represent the budget shares and the expenditure
elasticity of products i, the uncompensated and compensated own-prices and cross-prices
elasticities between products i and j. The elasticities of the CDE demand system must
imperatively satisfy a set of microeconomic conditions13:
• The Engel aggregation condition:
∑i
wiηi = 1 (20)
• The Cournot condition: ∑i
wiσuij = −wj (21)
• The Slutsky’s equation:
σcij = σuij + ηiwj (22)
• The symmetry condition:
σuij =wjwiσuji + wj(ηj − ηi) ∀i 6= j (23)
To these microeconomic conditions, the condition∑
iwi = 1 is obviously imposed.
3 Method, data and implementation
To use these two specifications CD and CDE, in applied models, the parameters of the
demand system must be estimated or calibrated. In the context of microsimulated model, the
CD demand system does not require any econometric estimates. Its parameters are, in fact,
budget shares and can be directly calibrated using a household survey. On the other hand, in
13To establish the complete conditions of the system (24), we relied on Bieri and Janvry (1972), Hanoch(1975), Aasness (1990), Deaton and Muellbauer (1980b), Tarr (1990), Sadoulet and De Janvry (1995) andChen (2017).
12
our approach, calibrating the elasticities and parameters of a CDE demand system requires
going through the parameter and elasticity estimation of the QUAIDS demand system as
we proposed in the previous section. In the following section, we detail the implementation
of that approach.
3.1 From QUAIDS estimation to CDE calibration
To implement our approach, we proceed as follows. In a first step, we estimate the parameters
λi, αi, βi and γij of a QUAIDS demand system, from the data of a household survey with a
numerically supported aggregation of products. The resulting estimates are used to compute
the different elasticities, always for the QUAIDS system.14
At the end of this step, we retain the QUAIDS estimates of the elasticities ηi, σcij and σuij
that become the initial values for the calibration of the elasticities and the parameters of the
CDE demand system. In all these steps, it is necessary to respect both the microeconomic
and econometric properties of the two demand systems (QUAIDS and CDE).
To do that, Chen (2017) proposes two calibration methods. The first one is a three-step
sequential type. It consists of calibrating in a sequence αi, ei and βi, from the elasticities
estimated in the previous step for a QUAIDS type demand system.15 The second method
is a kind of entropy method for which the calibrated values of the substitution (αi) and
expansion (ei) parameters are obtained simultaneously by optimizing an objective function
(or criterion) with the estimation obtained always in the previous step, for these elasticities
as initial values.
In this article, we improve and systematize the method proposed by Chen (2017). For this
purpose, we draw on the approach proposed by Golan et al. (1994) who have proposed an
optimization method to find unknown values of a matrix from economic data and respecting
many constraints or conditions.
14There are at least two STATA commands to estimate the parameters of a QUAIDS system. The firstone (quaids) was developed by Poi (2002) and updated in Poi (2012) and Poi (2013). The second command(aidsills) was developed by Lecocq and Robin (2015).
15The GAMS program of the sequential approach is in appendix of the article in Chen (2017).
13
Indeed, we apply the cross-entropy method based on Golan et al. (1994) by minimizing an
accurate criterion to deduce the calibrated values of the elasticities of a CDE-type function.
This criterion is based on the sum of the weighted ”differences” between the logarithms of
the unknown values of the elasticities (the CDE ones) and the values estimated from the
QUAIDS demand system.
The associated minimization program respects the definitions of the elasticities of the CDE
demand system (14, 15, 17, 18 and 19), the previously presented microeconomic constraints
(20, 21, 22 and 23) and those on the signs of the elasticities to be obtained. The solution
of this minimization program are the calibrated values of the parameters ei and αi and the
elasticities of the CDE demand system (ηi, σuij, σ
cij, σij). The complete minimization program
is presented in detail in System 24. This program is naturally solved for each household in
the survey. In our opinion, this is a major contribution in the literature on the calibration
of the parameters of a CDE demand system and in its use to approach the impact of public
policies in a microsimulation model.
14
Minαi,ei
[∑i ηi ln
(ηiηti
)]+[∑
i
∑j σ
cij ln
(σcij
σctij
)]+[∑
i
∑j σ
uij ln
(σuij
σutij
)]+[∑
i
∑j σij ln
(σijσtij
)]under constraints ∀i, j = 1, ..., k:
σij = αi + αj −∑
iwiαi −δijαi
wi
σuij = wj
[αj − ei(1−αi)∑
i wiei−
∑i wieiαi∑i wiei
]− δijαj where δij =
1 if i = j
0 else
ηi = (∑
iwiei)−1 [ei (1− αi) +
∑iwieiαi] + (αi −
∑iwiαi)
σcij = σuij + ηiwj∑iwiηi = 1∑iwiσ
uij = −wj
σuij =wj
wiσuji + wj(ηj − ηi) ∀i 6= j
(ηi − 1).(ηti − 1) > 0
ηi.ηti > 0
σcij.σctij > 0
(24)
where ηti , σutij and σctij are respectively the expenditure elasticity of products i, the uncom-
pensated and compensated own-prices and cross-prices elasticities between products i and
j, estimated on the QUAIDS. It is important to note that σtij does not come directly from
the QUAIDS system estimation. It is deduced in one step of the resolution for System 24 as
explained in Section 3.3.
To complete the calibration of the CDE demand system, the βi parameter still needs to be
calibrated. This one is obtained by calculation from the demand function of the good i in
the sequential approach. Following Chen (2017) and van der Mensbrugghe (2020), we use
the calibrated value of αi and normalizing u =1, p0i =1 and q0i = wi from (12) :
βi =wi
1− αi/∑i
wi1− αi
. (25)
15
In this article, the resolution of the system has been carried out using GAMS software after
importing the elasticities estimated from a QUAIDS function in previous steps from STATA.
There outputs are estimated values of αi, βi, ei, ηi, σij, σuij and σcij for a CDE demand system
for each household in the survey and products used in the estimation. Figure 1 summarizes
the structure of these steps.
3.2 Data
To estimate and calibrate the two considered demand systems, we need data from a household
survey on expenditures by product. We also need a vector of price indices for the products
considered, particurlaly to estimate the parameters of the QUAIDS demand system. We
present the two sources of data we used in this article.
3.2.1 Household survey
We use data from the 2019 wave of the Enquete Panel des Menages (EPM) of the Observa-
toire national du developpement humain (ONDH) of Morocco. This survey was implemented
for the first time in Morocco in 2012. In 2017 and 2019, the ONDH sought to ensure regional
representativeness (the 12 regions of the country in addition to national representativeness
and by place of residence) by increasing the sample size. The uncleared file we ran con-
tains 16,879 households. After clearing and discarding households with outliers or missing
observations, 15,904 households were considered in our application. Table 1 presents the
distribution of the ONDH wave 2019 EPM sample at the household level, as inferred from
the files used.
The adopted aggregation is based on the nomenclature used by the Direction de la Compt-
abilite Nationale (DCN) of the Haut-Commissariat au Plan (HCP). We mapped the 1,283
goods and services included in the survey (consumer expenditures) to four aggregated prod-
ucts for numerical and practical considerations. Indeed, firstly, the available STATA com-
mands (quaids and aidsills) used for estimating the parameters allow only a limited number
16
Figure 1: Structure of the three stages
Intput : budget shares, own and cross compensated and uncompensated price elasticities and expenditure elasticities estimated from a QUAIDS
demand system
(STATA output)
Output : calibrated parameters a, b and e; own and cross compensated and uncompensated price elasticities, AUES elasticities and expenditure
elasticities calibrated for a CDE demand system
(GAMS output)
Solved by cross entropy optimisation programs with GAMS
Source: Authors
17
Households Sample Freq. % Pop. Freq. Pop. %Urban 9 170 57.66% 4 841 027 66.29%Rural 6 734 42.34% 2 461 643 33.71%
Morocco 15 904 100.00% 7 302 670 100.00%
Source: Authors
Table 1: Household frequencies (sample and population) of the Survey EPM 2019 by placeof residence
of goods and services. Especially with a larger number of categories, the budget shares of
the products for different households are too low, which generates numerical problems at
different levels at all stages of the method. The selected aggregation into four product cate-
gories avoided this problem without affecting the object of our analysis.16
Secondly, Morocco has five levels of value added tax in 2019: 0%, 7%, 10%, 14% and 20%.
We have considered that the number of products taxed at 14% is low. For this reason, we
have chosen to aggregate these products with those taxed at 10%. Thus, all the analysis
conducted in this article is made on the basis of four value-added tax rates: 0%, 7%, 10%
and 20%. After mapping, each of the 1,283 products consumed or not consumed by each
household is placed in only one of these four categories. On the base of this mapping, we then
computed for each household the total expenditure for each of the four product categories
and computed the four associated budget shares.
3.2.2 Construction of the price vector
The basic price indices are derived from the HCP’s briefing note on the Consumer Price Index
(CPI) for the year 2019.17 Four price indices are computed as a weighted arithmetic mean
for each household, the weights are the budget shares computed from the 1,283 products.18
As mentioned earlier, to estimate the QUAIDS model on cross-section data, it is necessary
16It should be noted that this is not a limitation of the proposed method, which remains valid regardlessof the number of categories of products retained. The problem encountered is then purely numerical.
17See https://www.hcp.ma/L-indice-des-prix-a-la-consommation-IPC-de-l-annee-2019_a2451.
html.18We used a preliminary aggregation into 39 products selected from an Input-Output table of the HCP.
18
to have variance in prices across households. Ideally, we would like to have prices of goods by
household or at least by region/area/province via other sources than the survey. Given the
available information for the Moroccan case, prices were standardized by place of residence
(urban - rural) to introduce some heterogeneity between households. In fact, since the group
of products with zero VAT rate contains mainly agricultural products and since these ones
are generally more expensive in urban area, we have considered that the index of these
products in rural area is five points of percentage lower than its level in urban area. We
consider the opposite for the other three price indices. In doing so, the problem of price
endogeneity and the use of the instrumental variable method is avoided.
Obviously, by doing so, we expect to have collinearity between the different price indices
of the group of products. After verification, the correlation matrix shows that the linear
correlation coefficients are all in absolute value, higher than 74%. This could naturally cause
a problem in terms of inference but not for the parameters and elasticities estimation that
concern us in this article.
3.3 From theory to practice: steps for implementation
The complexity of the approach developed and the applied contribution of this article, con-
cerns the numerical implementation of the link between the estimation of the QUAIDS
demand system and the calibration of the one resulting from the CDE specification. In-
deed, from the modeler’s point of view, the microeconomic constraints that must be satisfied
are not necessarily part of econometricians’ concerns in the estimation process. In order to
ensure that we have microeconomic consistency in the estimation and calibration of expen-
ditures, own-prices and cross-prices elasticities and Allen-Uzawa elasticities of substitution,
we have broken down the work into six (6) main steps. One of them was done in STATA
(quaids estimation) while the other five were done with the GAMS software (calibration
steps, calculation of the β parameter). Flowchart 3 in Annex A summarizes each of the
steps we are going to detail in this section.
19
1. In the first step, after we have prepared the database according to the approach ex-
plained in the section 3.2, we estimate the QUAIDS demand system using the STATA
command of Poi (2012). The estimate is performed decile by decile in each residence
area (20 estimation procedures). Furthermore, we estimated the model retaining three
demographic variables: the age and the sex of the head of household and the size of
the household.
We also tested whether the form of the expenditure system was AIDS or QUAIDS
(H0 : λ = 0) for each decile. We conclude that the demand system is a QUAIDS type
for all deciles in the both areas. It is therefore the estimates of the elasticities of a
QUAIDS demand system obtained at this stage that will be used as a starting point for
initialization for the subsequent stages. Thus, at the end of this first step, we obtain
the estimates of the elasticities of a QUAIDS-type demand system for four aggregeted
products and for each household, namely σut1ij , σct1ij and ηt1i .
2. This step consists of importing the results obtained in STATA in step 1 into GAMS,
namly matrices of two (household, product i for ηti) and three dimensions (household,
product i, product j for σutij , σctij ). Before proceeding with the subsequent calibration
steps, we verified the microeconomic consistency, notably the respect of four conditions
(20, 21, 22 and 23) as well as the sum of the budget shares evaluated for each household
equal to one.
3. In this third step, we adjust the variables wi, ηt1i , σut1ij and σct1ij for each product i and
each household obtained in STATA, that do not eventually respect the microeconomic
constraints by performing a first calibration step. The objective is to obtain adjusted
shares using an entropy criterion so that the microeconomic consistency is respected
and re-evaluate all the other variables. The initial values used in this step are those
obtained at the end of the previous one. Thus, at the end of this step we have adjusted
shares (w2i ) consistent and that we will use in all future steps without any modification.
20
The solution of all other the variables of interest are then re-evaluated and will serve
as a starting point for the next step. We note them ηt2i , σut2ij and σct2ij .19
4. In this step, we start the calibration of the elasticities and parameters of the CDE
demand system. Given the non-linearity and complexity of the system 24, numerical
resolution turned to be difficult. The problem encountered is mainly related to the size
(number of constraints to be respected and variables considered) rather than to the
specification of the system itself. We have therefore chosen to follow Chen’s (Chen,
2017) example by adopting a two-step sequential approach.
• From the equation 19, the expenditure elasticity (η) depends only on the α and e
parameters.20 In a first step, we calibrate η for each product and for each house-
hold using an entropy criterion relative only to η and respecting the constraints
that only show η (i.e. 22, 23, Engel and sign conditions) and the budget shares
obtained in step 3. At the end of this step, we obtain a consistent estimate of η
and first estimates for the α and e parameters of the CDE demand system. In this
step, we use as an initial point the results obtained in step 3 i.e. w2i , η
t2i , σut2ij and
σct2ij . The results obtained are denoted ηt3i , σut3ij , σct3ij , αt3i and et3i . Furthermore,
with the values obtained at this step, we compute for the first time an estimate
of the Allen-Uzawa elasticities compatible with the CDE system (equation 13)
noted σt3ij .
• Then, we calibrate the compensed and uncompensated own and cross-prices elas-
ticities and the Allen-Uzawa elasticities of substitution using a new entropy cri-
terion and the constraints associated with prices elasticities (Cournot, 22, 23 and
the sign condition) and taking the values (w2i , σ
t3ij , σ
ut3ij , σct3ij and ηt3i ) as an initial
point.
19Note that at the end of this step, we have solved the estimation-calibration link per household for aQUAIDS-type application system that can be used in a microsimulated model.
20We proceed to the inverse of Chen (2017) which starts by calibrating the elasticity of the expenditurewhich depends only on the single parameter α.
21
At the end of this step, we obtain the new calibrated values of compensed and uncom-
pensated own and cross-prices elasticities and the Allen-Uzawa elasticities of substi-
tution σct4ij , σut4ij , σt4ij and we re-evaluate the expenditure elasticities, denoted ηt4i and
the CDE parameters, αt4i and et4i . These values are used in the next step as a starting
point for the last calibration step.
5. This last calibration step in our approach consists in applying an entropy criterion for
all the parameters and elasticities of System 24 using as an initial point the values
obtained in the previous steps. At the end of this step, we obtain the final parameters
and elasticities for the CDE demand system, calibrated for four products and for each
household.We denoted these by ηtF inali , σtF inalij , σutF inalij , σctF inalij , αFinali and eFinali .
6. Following Chen (2017) and van der Mensbrugghe (2020), we calibrate βi by a calcula-
tion directly from equation 25 and using the αFinali obtained in the previous step. For
each household, βi has been calibrated by normalizing u = 1, pi0 = 1 and qi0 = w2i .
21,22
At the end of this step, we get all the parameters and elasticities of a CDE-type demand
system for each household. In the following section, we use these estimates to compare the
results of simulations of changes in value added taxation in Morocco by considering the two
specifications CD and CDE presented previously into a microsimulated partial equilibrium
model.
4 Simulations of fiscal policies
To carry out the comparative analysis, we develop a partial equilibrium model that holds in
parallel the two demand systems showing the variables for the reference situation and for all
21Note that theoretically the calibrated βi can be positive or negative depending on the product and thevalue of the αi (especially when αi > 1). In our case, to respect the conditions on the parameters derivedby Hanoch (1975), we have upper bounded αi to 0.99 so that βi is positive.
22See for example Yu et al. (2004), footnote 10 of page 18.
22
the simulations considered. We begin by describing the model followed by the presentation
of the simulated scenarios and finally the results obtained.
4.1 Partial equilibrium model
We consider a partial equilibrium model in which the prices of different goods and total
expenditures are exogenous for each household. The prices of goods incorporate different
VAT rates by product group i, i = 1, ..., k = 4 as explained above. These rates are subject
to exogenous variations that we will illustrate in four VAT reforms. In parallel, we consider
three levels of expenditures per household, namely the statu quo compared to the reference
situation, a uniform increase of 5% and a decrease of 5%. Overall, we simulate twelve (12)
scenarios indexed (s) and which we will detail in the following sub-section.
In the model of Table 2, the variables in dot represent relative changes and those indexed 0
are relative to the baseline situation. The only new variable in the model is the price index
specific to each household, P .23 It allows to adjust the expenditure for each scenario (msa) to
make it comparable with the reference one and a constant poverty line. It is these variables
that are used in the distributional analysis.
Following Corong et al. (2017) and van der Mensbrugghe (2020), the equation 26 is the
relative change in the quantity demanded of each good i as a result of the relative changes
in prices and total expenditure in a CDE demand system:
qsi = ηims +∑j
σuij psj . (26)
It is interesting to note that in terms of relative variations, the CD demand system is a
special case of the CDE demand system. The equation 26 simply becomes for the CD
demand system qsi = ms − psi since in this case, ηi = 1, ∀i = 1, ..., k et σuij = 0 si i 6= j et
23In this article, we have retained two formulations of this price index. The multiplicative form, presented
in table 2, and an additive form PAs =∑k
i=1 wsi
(psi
p0i
)as proposed by Corong et al. (2017) and van der
Mensbrugghe (2020). The results obtained are approximately the same.
23
σuii = −1.
CD CDEqsi = ms − psi qsi = ηim
s +∑
j σuij p
sj
wsi =psi q
si
ms = α0i wsi =
psi qsi
ms
psi =dpsipsi
msi =
dmsi
msi
ms = m0(1 + ms)psi = p0i (1 + psi )
P s =∏k
i=1
(psip0i
)wsi
msa = ms
P s
Sources: Authors
Table 2: Partial equilibrium model
This model has been used to perform the twelve simulations presented below.
4.2 Presentation of scenarios
During the 1980s, Morocco undertook a major reform of its fiscal system. Its objective
was to develop a modern, coherent, efficient, and more universal fiscal system. Since then,
several measures have been introduced through successive finance laws. Value added tax
(VAT) is a major indirect tax that came into force in Morocco in 1986. It generates the
most tax revenue for the government, affecting both domestic and imported products. VAT
is defined on the basis of consumption and its coverage is very broad. Nevertheless, some
sectors remain outside this taxation, especially the agricultural sector. Some retail sales and
services or products are exempt by law. In 2019, the VAT rates applied in Morocco are 0%
for basic goods, 7%, 10%, 14%, and 20%. It is proportional and affects the entire population
in the same way.
In May 2019, the Third National Assises of Fiscality were held. The two previous ones being
held in 1999 and 2013. At each of these Assises, the reform of VAT was on the agenda,
in particular to make it a coherent, sustainable and neutral tax. Among the recommenda-
tions, it is considered to introduce a reduced rate on consumer goods that could currently
24
be exempt, including agricultural products. Another plans to consider only two rates, an
intermediate rate (10 - 12%) and a standard rate of 20%.
In this article, we simulate the impact on poverty of twelve scenarios by considering four
changes in VAT rates that we combine with three changes in the level of household expendi-
ture. In three cases, noted a, b, and c, we consider that the household expenditures remains
constant with respect to the baseline situation. In the other two cases, we simulate an in-
crease and a decrease of 5% in these expenditures respectively. For each of these three cases,
we apply four changes in the VAT rate, noted as 1, 2, 3 and 4:
• Case 1: Application of three rates namely 7% for products currently exempted or taxed
at 7% ; 14% for products taxed at 10% and 14% and 20% for other goods and services.
• Case 2: Application of a flat rate of 12% for all categories of goods and services.
• Case 3: Application of two rates, i.e. 10% for products currently exempt or taxed at
7%, 10% or 14% and 20% for other goods and services.
• Case 4: Application of a flat rate of 20% for all categories of goods and services.
4.3 Tools for distributional analysis
To analyze the impact of VAT rate reforms, combined or not with changes in per capita
household expenditures on poverty, we adopte two complementary approaches, a point esti-
mation and distributional analyse. Comparisons are made on resuslts obtained from model
based on a CDE-type demand system and a CD-type system.
For the point estimation analysis in terms of monetary poverty, it is necessary to set a
poverty line (noted zp). Since the distributions of per capita household expenditures have
all been adjusted to be comparable with the reference distribution, the zp line is the same
regardless of the distributions.
25
Scenarios Simulations conductedSim a-1 Fixed expenditures and 3 VAT rates (7%, 14% et 20%)Sim a-2 Fixed expenditures and flat rate (12%)Sim a-3 Fixed expenditures and 2 VAT rates (10% et 20%)Sim a-4 Fixed expenditures and flat rate (20%)
Sim b-1 Expenditures - 5% and 3 VAT rates (7%, 14% et 20%)Sim b-2 Expenditures - 5% and flat rate (12%)Sim b-3 Expenditures - 5% and 2 VAT rates (10% et 20%)Sim b-4 Expenditures - 5% and flat rate (20%)
Sim c-1 Expenditures + 5% and 3 VAT rates (7%, 14% et 20%)Sim c-2 Expenditures + 5% and flat rate (12%)Sim c-3 Expenditures + 5% and 2 VAT rates (10% et 20%)Sim c-4 Expenditures + 5% and flat rate (20%)
Sources: Authors
Table 3: Resume of scenarios
In this article, we choose a relative poverty lines. These poverty lines (distinct from one mi-
lieu to another) are computed on the distribution of per capita expenditures in households
in the reference situation. More explicitly, we use a relative poverty line equal to half of
the median of this distribution (zp = 0.5 median) according to place of residence (Ravallion
et al., 1994).
Using these poverty lines, we computed for the reference situation and for the twelve (12)
scenarios all the usual measures of monetary poverty (FGT) and inequality (Gini) at the
national level and for the two aeras of residence (urban and rural).24
In addition, and in order to go beyond the setting of a poverty line in the comparisons, we
have carry out a stochastic dominance analysis. Indeed, under the scenarios considered we
anticipate shifts in the distributions curves. Also, first (or second) order stochastic domi-
nance analysis is more informative in this context. Indeed, with these stochastic dominance
24Given the objective of this work, we present and analyze in this article only the variation for thepoverty incidence at the national level. Results by area of residence and for depth and severity of povertyand inequality measures are not presented but are available from the authors.
26
tests, the comparisons are unambiguous for a wider range of poverty lines.
4.4 Results
Our analysis highlights several interesting results depending on the focus chosen, namely the
economic policy dimension (tax reform and/or expenditure variation) and the methodolog-
ical dimension that allows for a comparison of the results according to the demand system
adopted. In order to conduct this comparison, we retain only what is related to the poverty
incidence.
4.4.1 Implications of a Tax Reform in Morocco
Looking only at the results obtained with the CDE demand system, it appears that without
a change in household expenditures, all of the proposed tax reforms will lead to an increase in
the poverty rates. Its incidence rises from 7.06% under the baseline situation to 8.01%, 8.57%,
8.71% and 10.99% according to the selected scenario. This implies that these reforms could
increase poverty. This can be explained by the fact that the products consumed mostly by
poor people are subject to an increase in the VAT rate, which reduces their purchasing power.
The simulations show that the increase in the poverty incidence is statistically significant for
all the proposed tax reforms. Indeed, none of the four confidence intervals contains the value
of the poverty incidence in the baseline situation (i.e. 7.06%). For the first scenario, which
consists of introducing a 7% VAT on exempt products, the poverty incidence increases from
7.06% to 8.01% . In the Scenario 4, which puts the rates at 20% for all products including
those currently exempt, the poverty incidence increases by 55.67% .
In terms of economic policy, it therefore appears that any tax reform designed to remove
tax exemptions for basic consumption goods, especially consumed by the poorest households,
must be done gradually so as not to have a major impact on these households. More generally,
this result highlights the social implications of any tax reform.
When the tax reform scenarios are combined with a negative shock that reduces household
27
expenditures by 5% (pandemic, drought, etc.) (Sim b − .), the poverty incidence increases
significantly under all four cases with, as before, stronger effects for cases 2 to 4. An opposite
effect is observed for case 1 when household expenditures benefit from a positive shock that
increases them by 5%. Indeed, for this reform, the poverty incidence is reduced compared
to its level in the baseline situation (6.57%). However, even with this uniform positive
growth (+5%) in household expenditures, under the other three cases, the poverty incidence
increases, but this increase is only statistically significant when all products are taxed at
20%.
4.4.2 CDE versus CD: is the calibration effort justified?
Table 4 has been constructed in order to compare the results of the simulations considered
under each of the two systems: the first, the CDE system, supposed to be flexible but chal-
lenging in terms of estimation and calibration, and the second, the CD system, constraining
but simpler to implement.
Several interesting results emerge from the analysis. For the 12 simulations considered, the
poverty incidence is systematically higher under the CD system. This result can be ex-
plained by the crucial role of the values of the elasticities in this type of analysis. Indeed,
with the CDE system the elasticities are different from one household to another, for each of
the groups of products considered (microsimulation effect) as opposed to the CD system for
which all the expenditure elasticities are constant and equal to the unit for all households
and for all products.25
More generally, the results of the simulations in these models reflect the expenditure elastici-
ties, own and cross-price elasticities estimated or calibrated for the selected demand systems.
In our case, household behaviour would probably be better approached with a CDE system
that captures more substitution between products and reduce the negative effects of the
25Indeed, our results show that 17.66% of households have an expenditure elasticity greater than one forproduct group 1. For the other three product groups, these proportions are 12.48%, 31.78% and 86.14%respectively.
28
impacts of the simulated shocks.
Incidence Lower bound Upper boundReference 7.06 6.57 7.55
No Growth (a)
Case 1CDE 8.01 7.49 8.53CD 8.13 7.60 8.65
Case 2CDE 8.57 8.04 9.11CD 8.86 8.31 9.40
Case 3CDE 8.71 8.17 9.25CD 8.85 8.31 9.40
Case 4CDE 10.99 10.39 11.60CD 11.80 11.18 12.42
5% decrease in household expenditures (b)
Case 1CDE 9.89 9.32 10.47CD 9.96 9.38 10.54
Case 2CDE 10.72 10.13 11.32CD 10.99 10.39 11.60
Case 3CDE 10.85 10.25 11.45CD 11.09 10.48 11.70
Case 4CDE 13.35 12.69 14.02CD 14.19 13.52 14.87
5% increase in household expenditures (c)
Case 1CDE 6.57 6.10 7.05CD 6.61 6.14 7.09
Case 2CDE 7.11 6.61 7.59CD 7.31 6.81 7.81
Case 3CDE 7.20 6.70 7.69CD 7.36 6.86 7.86
Case 4CDE 8.91 8.36 9.46CD 9.57 9.00 10.14
Sources: Authors
Table 4: Poverty incidence (%) and confidence intervals
The results show that the point estimates of the poverty incidence under the CDE system in
three simulations conducted (Cases 1 to 3), are each time included in the confidence intervals
constructed for the same measure under the CDE system and vice versa. Tests of equality
between these two measures in each of the cases considered, lead to the non-rejection of the
29
underlying null hypotheses. This means that the differences between the two systems are not
statistically significant. However, when the shock is greater (Case 4), the differences become
statistically significant and the point estimates of the poverty incidence fall outside the
confidence intervals of the alternative demand system. For example, in simulation Sim b-4,
the estimated poverty incidence is equal to 13.35% under the CDE system and 14.19% under
the CD system. The respective confidence intervals range from 12.69% to 14.02% under the
CDE system and from 13.52% to 14.87% under the CD system. This underlines the fact
that as soon as the simulated shocks are large, it is probably preferable to retain the CDE
system. In the contrary case, the precision gain does not seem to justify the use of the CDE
system in terms of point estimation regarding the cost in terms of estimation-calibration.
Figure 2: First-order stochastic dominance Sim-b-4
Cum
ulat
ive
perc
enta
ge o
f hou
seho
lds
4000 4600 5200 5800 6400 7000
Household per capita annual expenditure
Exp.PerCap - BaselineExp.PerCap - CDESimb_4
Exp.PerCap - CDSimb_4
Source: Authors
All of these results could be conditioned by the fact that they are associated with an inci-
dence measure based on a poverty line. We have therefore considered a stochastic dominance
30
robustness analysis. For this purpose, the adjusted household per capita expenditure den-
sities and distribution curves obtained from the two demand systems and compared to the
baseline situation have been constructed.
Figure 2 illustrates, for example, the distribution functions of the three per capita expen-
ditures of the Sim b-4 simulation featuring a uniform 5% decrease in expenditures and a
20% VAT rates for all products. It appears that whatever the value of the poverty line be-
low MAD 7, 000 per person and per year, poverty increases under the two demand systems
considered. Moreover, the gap between the curves increases as the value of the poverty line
increases and the gap between the simulated curves (CDE versus CD) becomes wider. This
shows that the results obtained under the CD system amplify the negative impact than those
obtained with a CDE system.
5 Conclusion and discussion
The objective of this article is to compare two demand systems, namely the CD-type system
and a more flexible but more complex to estimate, the CDE demand system, in a microsim-
ulation model. For this purpose, we developed a hybrid estimation-calibration approach to
calibrate the parameters and elasticities of the CDE demand system using estimates obtained
from the QUAIDS demand system and numerically minimizing an entropy criterion.
Using Moroccan ONDH data (wave 2019 of the Panel Survey), we construct a partial equilib-
rium microsimulation model to simulate four cases of VAT reforms with or without changes
in household expenditures on the usual monetary poverty measures. We then perform point
comparisons, confidence intervals and stochastic dominance analysis on the poverty measures
between the two systems (CDE, CD).
It appears that when the simulated shocks are moderate, the gain of using a CDE system
instead of a CD system is low. Estimates of poverty measures are statistically equivalent. On
the contrary, when the simulated shocks are stronger, the differences become significant and
31
increase as the poverty line increases when the analysis is performed in terms of stochastic
dominance.
At the theoretical level, our results show that it is clearly different to use one or other de-
mand systems given their characteristics (parameters and elasticities). In practice, the use of
these models can lead to different results when evaluating public policies and their impacts
on poverty measures.
However, our conclusions are obviously conditioned by the context considered. They are
inferred from aggregation and specific simulations in the framework of a partial equilibrium
model in which the interactions between different prices on the one hand and between prices
and demands on the other are not taken into account (no general equilibrium effects). The
differences obtained could indeed be greater or smaller in different contexts.
Based on the estimation-calibration approach of a CDE demand system developed in this
article, we have been able to highlight by comparison, several relevant results from both
theoretical and empirical points of view. These results were not trivial a priori.
The approach developed in this article is numerically time-consuming and demanding. If
we want our conclusions to be taken into account in a public policy simulation exercise, it
is important to make it systematic by automatizing its numerical implementation. In this
way, modellers and decision-makers will be better supported in their efforts. It is precisely
in this sense that we are pursuing our research.
32
References
Aasness, J. (1990). Properties of demand functions for linear consumption aggregates. Tech-nical report, Discussion Papers.
Abdelkhalek, T. and Dufour, J.-M. (1998). Statistical inference for computable generalequilibrium models, with application to a model of the moroccan economy. Review ofEconomics and Statistics, 80(4):520–534.
Attfield, C. L. (1985). Homogeneity and endogeneity in systems of demand equations. Jour-nal of Econometrics, 27(2):197–209.
Banks, J., Blundell, R., and Lewbel, A. (1997). Quadratic engel curves and consumerdemand. Review of Economics and statistics, 79(4):527–539.
Bieri, J. and Janvry, A. d. (1972). Empirical analysis of demand under consumer budgeting.Technical report.
Boeters, S., Bohringer, C., Buttner, T., and Kraus, M. (2010). Economic effects of vatreforms in germany. Applied Economics, 42(17):2165–2182.
Chen, Y.-H. H. (2017). The calibration and performance of a non-homothetic cde demandsystem for cge models. Journal of Global Economic Analysis, 2(1):166–214.
Corong, E., T., Hertel, R., McDougall, M., T., and van der Mensbrugghe, D. (2017). Thestandard gtap model, version 7. Journal of Global Economic Analysis, 2(1):1–119.
De Quatrebarbes, C., Boccanfuso, D., and Savard, L. (2016). Beyond representative house-holds: The macro–micro impact analysis of vat designs applied to niger. Economic Mod-elling, 57:76–92.
Deaton, A. (1988). Quality, quantity, and spatial variation of price. The American EconomicReview, pages 418–430.
Deaton, A. and Muellbauer, J. (1980a). An almost ideal demand system. The Americaneconomic review, 70(3):312–326.
Deaton, A. and Muellbauer, J. (1980b). Economics and Consumer Behavior. CambridgeUniversity Press.
Dimaranan, B., McDougall, R., and Hertel, T. (2006). Behavioral parameters. Global trade,assistance, and production: the GTAP, 6:1–17.
Golan, A., Judge, G., and Robinson, S. (1994). Recovering information from incompleteor partial multisectoral economic data. The Review of Economics and Statistics, pages541–549.
Hanoch, G. (1975). Production and demand models with direct or indirect implicit additivity.Econometrica: Journal of the Econometric Society, pages 395–419.
Lecocq, S. and Robin, J.-M. (2015). Estimating almost-ideal demand systems with endoge-nous regressors. The Stata Journal, 15(2):554–573.
Ndemezo, E. and Baye, F. M. (2016). Evaluation of redistributive and welfare impacts ofindirect taxes reform in rwanda. In Poverty and Well-Being in East Africa, pages 165–188.Springer.
Poi, B. P. (2002). From the help desk: Demand system estimation. The Stata Journal,2(4):403–410.
Poi, B. P. (2012). Easy demand-system estimation with quaids. The Stata Journal,12(3):433–446.
Poi, B. P. (2013). Software updates. The Stata Journal, 13(3):667.
33
Ravallion, M. et al. (1994). Poverty comparisons, volume 56. Taylor & Francis.Ray, R. (1983). Measuring the costs of children: an alternative approach. Journal of Public
Economics, 22(1):89–102.Rimmer, M. T. and Powell, A. A. (1996). An implicitly additive demand system. Applied
Economics, 28(12):1613–1622.Sadoulet, E. and De Janvry, A. (1995). Quantitative Development Policy Analysis. A Johns
Hopkins paperback. Johns Hopkins University Press.Savard, L. (2004). Un systeme de demandes aids dans un contexte egc micro-simulation pour
l’analyse de pauvrete et des inegalites. Cahier de recherche/Working Paper, 4:10.Shoven, J. B. and Whalley, J. (1992). Applying general equilibrium. Cambridge university
press.Surry, Y. (1993). The constant difference of elasticities function with applications to the ec
animal feed sector. Journal of Agricultural Economics, 44(1):110–125.Tarr, D. G. (1990). A modified cournot aggregation condition for obtaining estimates of
cross-elasticities of demand. Eastern Economic Journal, 16(3):257–264.van der Mensbrugghe, D. (2020). Mitigation, adaptation and new technologies applied gen-
eral equilibrium (manage) model: Version 2. Model documentation available upon request.Yu, W., Hertel, T. W., Preckel, P. V., and Eales, J. S. (2004). Projecting world food demand
using alternative demand systems. Economic Modelling, 21(1):99–129.
34
Appendix A. Flowchart of implementation
Figure 3: Implementation steps
STEP 1: Estimation of the QUAIDS parameters + elasticities (expenditures, own-prices and cross-prices
compensated and uncompensated) for 4 productsSTATA: command quaids evaluated for each household
Input: pi, wi, m and demographicOutput: ht1, sc
ijt1, su
ijt1
STEP 2: Reading results in GAMSVerification: the microeconomic constraints associated with the
demand systems are strictly respected at the end of step 1.
STEP 3: Numerical adjustment of variables to respect the micro-economic consistency of
the four system constraints (Cournot, Engel, Slutsky, symmetry) in GAMS
Input: ht1, scij
t1, suij
t1, wi
Output: ht2, scij
t2, suij
t2,wi2 consistent
STEP 4: Sequential calibration in GAMS4-1. Variable in the criterion: hi
From (19) and under Engel, Cournot, symmetry and Slutsky constraints and signs on hi + computation of
suij, sc
ij and sijt with a3
i e3i obtained at this step
Input: hit2, sc
ijt2, su
ijt2, wi
2
Output: a3i, e3
i, hit3, sij
t3, scij
t3, suij
t3
4-2. Variable in the criterion: suij, sc
ij and sijt from (13)
and (16) and under the following constraints : Slutsky, Cournot, symetry and signs + computation of hi with
a4i and e4
i obtained at this stepInput: a3
i, e3i, hi
t3, sijt3, sc
ijt3, su
ijt3, wi
2
Output: a4i, e4
i, hit4, sij
t4, scij
t4, suij
t4, wi2
STEP 5: Calibration of hi, suij, s
cij , sij of the
complete system (24) under all constraints (all variables in the criterion)
Input: a4i, e4
i, hit4, sij
t4, scij
t4, suij
t4, wi2
Output: eFinali,a
Finali,h
Final, sijFinal, sc
ijFinal, su
ijFinal,
wi2
STEP 6: Calculation of bi from (25)Intput: aFinal
i, wi2
Output: eFinali,a
Finali, bi
Final, hFinal, sijFinal, sc
ijFinal,
suij
Final, wi2
Estimation Calibration with relevant entropy criteria (cross entropy) Calculation
Source: Authors35